Segmenting Wildfire Burned Area in the 🌍

Data Sources

Our data is from cropped Landsat Collection 2 imagery for all available fires in California. These labels are extracted from the MTBS, a project that measures severity and burned areas across the US. Notice that each file name corresponds to the Event_ID in the MTBS dataset. The folder contains pre-image and post-images, this is corresponds to data before and after the fire event. Thus, you should see fire scars in all the post-images.

Both pre- and post- imagery is captures during the fire-season (between day 152 and 273 of the year in California). Since Landsat has an average revisit day between 8 and 16 days, many of the fires have several images. Notice that we have visible bands and infrared bands to calculate different vegetation and burning indexed, this might be relevant if you want to use this data for pre-training a model from scratch, but feel free to ignore it if you prefer fine-tune something. Native resolution for Landsat is 30 meters, which works pretty well for some of the purposes ecologists are interested in, but not for segmentation (a lot of the papers use Sentinel-2 data, 3 times higher resolution, for example), we should test this.

Environment

Results were obtained using python 3.9. The required packages can be installed with

pip install -r requirements.txt

In particular, we use pytorch lightning 2.1.0 which requires python >= 3.8 as well as pytorch = 1.12.1 with cuda 11.3. However, the training code 'should' work with any version of pytorch >= 1.10.0 and python >= 3.6.0 assuming pytorch lightning works. Although note that these requirements were ported from conda and may not be fully accurate. In particular, when downloading the data, gdal will not properly install with pip unless a built wheel is already made for your system. If this is not the case, we recommend using the provided environment.yaml to create a conda environment with all the required packages.

conda env create -f environment.yaml

Data

Raw data can be downloaded by setting the root path defined in the bash script and running it

bash data_dowload_submission/get_data.sh

To download data for the scaled experiments run (Note this can take a very long time due to amount of data)

bash data_dowload_submission/get_data_scaled.sh

Training a segmentation model

A unet model with resnet34 backbone can be trained by running

python train.py \
    --model_type "unet" \
    --backbone "resnet34" \
    --pretrained True \
    --device 0 \
    --num_workers 8 \
    --log_dir "~/logs" \
    --data_dir "~/data/processed" \
    --split 0.8 \
    --batch_size 32 \
    --crop_method "scale" \
    --crop_size 256 \
    --overlap 64 \
    --post_only False \
    --dnbr True

For a full set of available models and backbones see https://smp.readthedocs.io/en/latest/. Results are saved to log_dir and can be visualized with tensorboard

tensorboard --logdir ~/logs